I build machine learning systems that reach production and stay there.

I'm an AI researcher at the (Advanced Computing Research Laboratory, Moscow), where I built the first operational malaria early-warning system covering every province of Angola — a five-engine ensemble forecasting outbreaks six to eight weeks ahead. Behind the research sits seven years of production engineering: fintech platforms scaled past 5,000 active users, fraud-detection models running in live financial systems, and deep-learning training accelerated 3× with CUDA and TensorRT.

Nominated for Forbes Africa Lusophone "Under 30" · Invited speaker at CPHIA 2026, Addis Ababa

An MCP server that gives AI agents access to provincial malaria surveillance and short-horizon outbreak forecasting — with the guardrails that make model output safe for an agent to act on.

The harness caught two defects I would not have found by inspection: ensemble weights that lost to a naive baseline at long horizons, and 80% intervals with 90–95% empirical coverage. Both are documented in the README rather than quietly fixed.

Python · MCP SDK · CI across 3.10–3.12 · MIT

Operational forecasting across all 18 provinces under the administrative division in force during the study period (2000–2024), combining climate covariates with epidemiological-memory features.

K-means stratification resolved the provinces into three epidemiological strata and exposed a 2.8× burden disparity, providing an evidence base for differentiated resource allocation. Featured by international media in four languages.

graph LR
    S[Provincial surveillance<br/>2000–2024] --> F[Feature engineering]
    C[Climate covariates] --> F
    M[Epidemiological<br/>memory features] --> F
    F --> E{Five-engine ensemble}
    E --> B[Rolling-origin<br/>backtest]
    B --> G[Guardrails:<br/>horizon + history checks]
    G --> A[MCP server<br/>agent-facing tools]
    G --> H[Provincial health<br/>authorities]

    classDef data fill:#1f4e79,color:#fff,stroke:none;
    classDef model fill:#276DC3,color:#fff,stroke:none;
    classDef out fill:#0b6b3a,color:#fff,stroke:none;
    class S,C,M data;
    class E,B model;
    class A,H out;

Measured internal performance against external generalisation failure, with Grad-CAM explainability to identify not just whether the model degraded but where its attention shifted when it did. Accuracy reported on internal validation is not evidence of clinical reliability elsewhere.

• Operational Malaria Forecasting in Angola Using Ensemble Models, Regional Clusters, and Epidemiological Memory Features — ResearchGate, Feb 2026
• A Hybrid Artificial Intelligence Framework for Extreme Pattern Discovery in Complex Systems — Article and Conference Paper, Sep 2025
• Internal Performance and External Generalization Failure of a Deep Learning Classifier for Mammographic Lesion Assessment: A Cross-Dataset Evaluation with Explainability Analysis — ResearchGate, Aug 2026

• Forbes Africa Lusophone "Under 30" Nominee (2026) — for AI innovation with social impact
• "Jovem da Diáspora que Honra Angola" — national distinction for diaspora achievement
• Letter of Recommendation, Artificial Intelligence Research Institute (AIRI) — Apr 2026
• Invited Speaker, 5th International Conference on Public Health in Africa (CPHIA 2026), Addis Ababa
• Speaker, National Forum on Artificial Intelligence (FNIA), Angola

Work covered by Forbes África Lusófona, Sputnik Africa, Pulse of Africa, Jornal de Angola, SAPO, PTI and Izvestia.

Generative AI & agents

ML & research

Backend & data

Cloud & DevOps

Security — BSc Information Security & Cybersecurity (in progress) · PWST (TCM Security) · PCI DSS (TÜV SÜD) · IAM/MFA in production

Turning the Claude ecosystem work into shipped code rather than certificates: an MCP server plus a RAG pipeline with an evaluation harness, in the epidemiological forecasting domain I know best. If a system is going to be consulted by an agent instead of a specialist, the guardrails have to travel with it.